GenAI trained for document processing: Advancing business operations at scale

We’re rapidly moving towards a world where the back office is almost fully automated. Consider: Gartner predicts that within a year, half of B2B invoices will be paid without manual intervention, further increasing to 80% by 2030.

We’re not there yet—recent research suggests that, today, 57% of invoice data still needs manual entry—but it’s clearly coming soon. That’s because the speed, capacity, cost savings, and accuracy delivered by modern intelligent document processing (IDP) are nearly impossible to resist. And adoption will accelerate even more as generative AI (GenAI) powers big leaps forward in IDP’s capabilities, robustness, and ease of implementation. 

UiPath is one of the major innovators in the IDP space. Our IDP innovation and product strategy can be summed up like this: leverage automation, specialized AI, and GenAI to blast through IDP roadblocks that historically have inhibited its adoption. These roadblocks include:

• Challenges in handling high variability in document formats

• Difficulty in handling unstructured data

• Laborious model training

• Scalability limitations

• Security concerns

Over the past year or so, we have made notable strides forward in addressing these issues. In fact, Everest Group recently named us, for the second year running, as a Leader in the IDP category PEAK Matrix® Assessment. This year, we had the highest ratings for both ‘vision and capability’ and ‘impact in the market.’ (You can register here to get an exclusive copy of the UiPath profile report.)

So, with that as the backdrop, let’s take a look at the ways we are combining automation, specialized AI, and GenAI to take IDP to the next level and make it an even more irresistible value proposition.

Using foundational LLMs to tackle unstructured data challenges

One of the ways we have pushed past existing boundaries in IDP is through our use of foundational LLMs like GPT-4, which are trained on diverse datasets.

We have incorporated new GenAI capabilities throughout our existing products. For example, we now use LLMs within UiPath Document Understanding, which has greatly improved our ability to accurately process freeform unstructured documents like legal agreements, contracts, and emails, at high scale. GenAI has also enabled us to dramatically reduce the effort it takes to train models to understand specific documents and forms—cutting the time up to 80%, from weeks to hours or a day or two.

At HUB International, a large insurance company, these new GenAI-powered capabilities have turned an impossible-to-do manual task into one that’s completed in just minutes:

We needed to pull and identify data from thousands of unstructured documents—a task that we haven't been able to complete manually ever before due to high complexity and variability of the documents. Now, generative AI capabilities in UiPath Document Understanding do that in less than five minutes—with incredible accuracy and time savings for our team.

Thach Nguyen, Director of Digital Innovation, HUB International

Shifting to specialized LLMs to increase accuracy and boost operational efficiency

While foundational LLMs have a broad task range they can carry out well, they have significant limitations and weaknesses when it comes to applying them to specific domains, business processes, or document types. This creates a problem in trying to leverage them for IDP (which is, at heart, all about processing specific document types within a particular, defined business context). 

Think of it: if you’re looking for an IDP solution, you probably don’t care if a model knows who directed Titanic or how photosynthesis works. Instead, you’d prefer a model that knows which fields are required to be filled in your company’s (or country’s) tax forms or which documents can be accepted as a proof of identity. You want a model that will have much higher accuracy right out of the box, require significantly less training time, and be far more computationally efficient. This is exactly what the right specialized LLMs can deliver.

That’s why we have been creating specialized LLMs built around a corpus of knowledge specific to domains, industries, and business processes. At the recent UiPath AI Summit, we announced the General Availability of two such models, DocPath and CommPath. While leveraging GenAI at the core, these models are specialized and tuned to process complex business documents and diverse communications—from intricate interactions to free-form documents and tables. 

By narrowing the focus, but retaining the vast new power of GenAI, our specialized LLMs can significantly outperform the output accuracy of currently available LLMs—which is critical in delivering efficient, trustable IDP. (Think of it: an extra zero extracted from an invoice can cause a payment to inflate tenfold or cause delays in accounts payable processes, potentially straining relationships with vendors, reducing efficiency, and even causing financial and reputational losses for the company.)

To check the accuracy of UiPath LLMs against both foundational LLMs and other IDP vendors, we performed head-to-head tests across a diverse set of enterprise documents. The results? Compared to leading GenAI models, DocPath error rates were 45% to 76% lower. Moreover, in interpreting complex table structures, UiPath DocPath outperformed other IDP and GenAI vendors by registering 30-65% fewer extraction errors.

Source: Test by UiPath AI R&D on a diverse set of enterprise documents

Specialized LLMs also prove beneficial in minimizing or even eliminating the need for custom training, as they are better tailored to handle various types of documents and communications out of the box. For example, the model that powers our free automated copy-and-paste standalone app, UiPath Clipboard AI™, delivers over 90% accuracy from the get-go for a range of frequently used documents such as invoices, utility bills, purchase orders, receipts, passports, and W2 forms.

Generative LLMs that are specialized for certain tasks enable enterprise automation at scale. For IDP, specialized LLMs that deliver lower error rates, improved performance, and accelerated time to value effectively transform unstructured and semi-structured documents and communications into valuable digitized data that can be used in automated processes—driving up an organization’s operational efficiency.

Ensuring consistency and security for GenAI deployments

Although the use of GenAI radically improves IDP’s performance and broadens its capabilities, it also introduces increased security and operational risks. For example, foundational LLMs require broad access to data to function effectively, so sensitive business information may be inadvertently exposed or mishandled. Moreover, foundational LLMs may lack predictable consistency, meaning they respond differently to the same questions at different times. Relatedly, they can generate so-called hallucinations as output—meaning that the model provides an inaccurate answer that is not based on any facts or logic, yet treats it as accurate.

To ensure the full power of GenAI can be released safely throughout the enterprise, these risks can and must be mitigated. We have taken several actions to do so.

To address the data security issues, UiPath provides an additional layer of security to generative models—the UiPath AI Trust Layer—which helps ensure adherence to compliance standards. Additionally, to mitigate the risk of errors, there are a few ways to validate the model output and keep a human in the loop to double-check on the results, as needed.  

Moreover, using specialized LLMs instead of foundational LLMs is, in itself, a substantial risk-reducer. For example, DocPath and CommPath process data securely within the secure environment of the UiPath cloud ecosystem. This not only helps ensure comprehensive adherence to compliance standards but also to maintain data privacy, significantly lowering the risk of confidentiality breaches. What’s more, DocPath and CommPath are far less likely to hallucinate, because their answers are based solely on the data and context from the company documents and messages they process.

LLMs at the forefront of digital evolution in IDP 

The emergence of GenAI within IDP has facilitated immense progress in business operations, yet these advancements come with their own set of challenges. Specialized LLMs play a vital role in overcoming these hurdles while improving operational efficiency. With their capability to intelligently process large amounts of unstructured data, LLMs are spearheading a new wave of productivity and growth in the business world. As the world continues to evolve digitally, their pivotal role in seamless data management will only become more pronounced.

UiPath has embraced the new possibilities emerging from GenAI and LLMs, using these advances broadly in our quest to deliver the best-in-class IDP and enterprise automation across various business processes. You can watch UiPath AI Summit to learn about the latest UiPath AI capabilities such as DocPath and CommPath LLMs.